RPANet: A Residual Pyramid Semantic Segmentation Network Based on Atrous Convolution

Semantic segmentation is a challenging task to achieve accurate pixel classification. Recent studies witnessed that residual networks can better solve problems such as gradient vanishing and exploding in neural network backpropagation. However, when objects are obstructed or certain target sizes are small, it can lead to insufficient feature extraction ability of the residual network. In this paper, we propose a residual pyramid network based on atrous convolution (RPANet) to obtain better semantic segmentation results and improve segmentation network efficiency. Specifically, we first put forward a novel residual pyramid architecture, which integrates contextual information with the relationship features between pixels, and outputs multi-level and different scale features. In addition, we propose multi-scale atrous convolutions, which can not only capture long-range contexts information and enlarge the effective receptive field, but also the size of the feature map is not reduced by the residual connection. Extensive experiments demonstrate the effectiveness and superiority of our method on Cityscapes and CamVid datasets. We demonstrate that our method can improve the semantic segmentation performance reduce network complexity.